ImageNet

Fei-Fei Li's premise, unfashionable in 2007, was that the bottleneck in computer vision was not the algorithms but the data, and that the way to make machines see was to show them the world at a scale no dataset had tried. ImageNet organised its images by the nouns in WordNet, tens of thousands of categories, and labelled them by paying workers on Amazon's Mechanical Turk, a service that was two years old. The version presented as a poster at CVPR in June 2009 had 3.2 million images; by 2010 it had 14 million.

The ImageNet Large Scale Visual Recognition Challenge began in 2010 with a thousand categories and 1.2 million training images. In 2010 and 2011 the winners used hand-designed features and support-vector machines, and error rates crept down by a point or two a year. In 2012 AlexNet cut the error by ten points at a stroke, and the field changed direction within months.

The dataset is the reason that happened when it did. Convolutional networks had existed since 1989; GPUs since 1999; what they lacked was a million labelled examples and a leaderboard on which winning was unambiguous. ImageNet supplied both, and made the benchmark, rather than the theorem, the field's unit of progress.